The $ refers to the address of the instruction itself, so the instruction is executed over and over again, until the device is no longer busy.

When I have a moment, I will summarize what I know. Actually, I learned something new by reading the Transcendent article, so thanks for that!

The thing to be concerned about is not that the routing will necessarily make the crystal go crazy, but that on one in a hundred boards, where by chance a poor sample of the crystal comes together with a chip at the edge of its specitication, the oscillator will sometimes stop when the battery is low. That's a hard problem to debug, and the reason for conservative design. Top of your list should be to add those decoupling caps!

Isn't CR2 the wrong way round for that?

The voltage doubler has the perhaps significant advantage that it won't be damaged if a software crash leaves the control signal stuck in either state.

I will have to get some inductors and try it out. Fewer components too!

Difficult to marry theory and practice in assessing the efficiency. From one point of view, every coulomb of charge that flows through the LED has been through one, two, three diode drops elsewhere in the circuit, so we would do well to beat 50%. From another angle, the current shown on my bench supply is about twice the average LED current as measured with a scope across a 10 ohm resistor in series with it. I've not yet had chance to measure the supply current more accurately than that.

@VoltageSpike: You may care to note that I have already been thanked by the kind person SadlerJ who asked the question, and that my answer does in fact address directly the two points they raised.

It's a wonderfully amusing circuit, and it's given us a lot of fun. I'm so glad we had enough stuff in the house to build it without venturing out for supplies!